Add the ability for callers to adjust explanatory variables to use when it cannot inverse a matrix

Author: usk83

v2/error.go

@@ -0,0 +1,74 @@
+package regression
+
+import (
+	"errors"
+	"math"
+
+	"gonum.org/v1/gonum/mat"
+)
+
+var (
+	ErrNearSingular    = &ConditionError{isExactlySingular: false}
+	ErrExactlySingular = &ConditionError{isExactlySingular: true}
+
+	matConditionError    = mat.Condition(0)           // matrix singular or near-singular
+	matConditionErrorInf = mat.Condition(math.Inf(1)) // matrix exactly singular
+)
+
+type ConditionError struct {
+	err               error
+	isExactlySingular bool
+	Hint              *ConditionErrorHint
+}
+
+func (e ConditionError) Error() string {
+	if e.err == nil {
+		return ""
+	}
+	return e.err.Error()
+}
+
+func (e ConditionError) Is(err error) bool {
+	if condErr, ok := err.(*ConditionError); ok {
+		return e.isExactlySingular == condErr.isExactlySingular
+	}
+	return false
+}
+
+func (e ConditionError) Unwrap() error {
+	return e.err
+}
+
+func wrapAsConditionError(err error, hint *ConditionErrorHint) *ConditionError {
+	return &ConditionError{
+		err:               err,
+		isExactlySingular: errors.Is(err, matConditionErrorInf),
+		Hint:              hint,
+	}
+}
+
+type ConditionErrorHint struct {
+	ExplanatoryVars []ExplanatoryVarHint
+}
+
+type ExplanatoryVarHint struct {
+	OriginalIndex int     // 元々のインデックス
+	Label         string  // 名称
+	Coeff         float64 // 偏回帰係数 B
+	VIF           float64 // 共線性の統計量 VIF
+}
+
+func newExplanatoryVarHints(basicRawModel *basicRawModel, vifs []float64) []ExplanatoryVarHint {
+	hints := make([]ExplanatoryVarHint, basicRawModel.numOfExplanatoryVars)
+
+	for i := 0; i < basicRawModel.numOfExplanatoryVars; i++ {
+		hints[i] = ExplanatoryVarHint{
+			OriginalIndex: basicRawModel.indexesTable[i],
+			Label:         basicRawModel.explanatoryVarsLabels[i],
+			Coeff:         basicRawModel.coeffs[i],
+			VIF:           vifs[i],
+		}
+	}
+
+	return hints
+}

v2/regression.go

@@ -2,6 +2,7 @@ package regression
 
 import (
 	"errors"
+	"fmt"
 	"math"
 	"strconv"
 
@@ -363,8 +364,15 @@ func (r *Regression) Run() (*Model, error) {
 
 	sRInv := new(mat.Dense)
 	if err := sRInv.Inverse(sR); err != nil {
-		logger.Err.Printf("Cannot inverse a matrix: %v", err)
-		return nil, err
+		e := fmt.Errorf("cannot inverse a matrix(sR): %w", err)
+
+		logger.Err.Println(e)
+
+		if errors.As(e, &matConditionError) {
+			return nil, wrapAsConditionError(e, &ConditionErrorHint{ExplanatoryVars: newExplanatoryVarHints(bm, coeffsVIFs)})
+		}
+
+		return nil, e
 	}
 
 	sInv := new(mat.Dense)
@@ -433,8 +441,15 @@ func (r *Regression) Run() (*Model, error) {
 	// 相関行列から偏相関行列を求める
 	corrDenseInv := new(mat.Dense)
 	if err := corrDenseInv.Inverse(corrDense); err != nil {
-		logger.Err.Printf("Cannot inverse a matrix: %v", err)
-		return nil, err
+		e := fmt.Errorf("cannot inverse a matrix(corrDense): %w", err)
+
+		logger.Err.Println(e)
+
+		if errors.As(e, &matConditionError) {
+			return nil, wrapAsConditionError(e, &ConditionErrorHint{ExplanatoryVars: newExplanatoryVarHints(bm, coeffsVIFs)})
+		}
+
+		return nil, e
 	}
 	coeffsPartialCorrelations := make([]float64, bm.numOfExplanatoryVars)
 	for i := range coeffsPartialCorrelations {